AI & Automation

Google Found an AI Agent Turning Patches Into Exploits, With No Human in the Loop

A Russian-speaking operation's exploit_agent framework compared web software before and after each fix, worked out the flaw, then wrote and ran the exploit by itself. For defenders, a published patch is now the start of the attack clock, not the end of the risk.

November 17, 2026 · AI & Automation
Close-up of a dark code editor screen at an angle, showing brightly coloured CSS rules for a slider container with position, width, margin and border-radius values

Key Takeaways

  • Google's Threat Intelligence Group says that in the past quarter it saw attackers use LLM frameworks to automate vulnerability discovery and exploitation without a human in the loop for the first time.
  • The framework, exploit_agent, runs continuous differential analysis of popular web software, comparing code before and after patch announcements, then drafts exploits and executes them automatically.
  • Exploited zero-days rose from a 2025 average of 8 a month to 11 a month in 2026, then spiked to 22 in August, while overall disclosures and exploits nearly doubled between January and August.
  • Google says these frameworks currently trade stealth for velocity, leaving forensic traffic, recognizable identifiers and code-level traces that make them easier to detect.

For years, the gap between a vendor shipping a security fix and attackers turning it into a working exploit has been the window that patch management lives in. Google now says that window is being worked by machines. In its latest threat report, the company describes a framework that watches popular web software for new patches, reads what changed, works out the flaw each fix was meant to close, and then writes and fires an exploit against real targets, all without a person directing any step. Google calls it a "critical inflection point."

What Google Found on a Staging Server

The evidence came from an exposed staging server. According to Google Threat Intelligence Group's AI Threat Tracker, published November 16, the server hosted a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, wired to several frontier and locally hosted language models. Google writes that it is set up to "conduct continuous differential analysis of popular web application software," comparing code before and after patch announcements, identifying likely vulnerabilities, drafting exploits and executing them "without human intervention." One figure in the report shows an agent analyzing an insecure access-control check in a WordPress page-builder plugin.

The operation was not a research toy. The same server held post-compromise scripts for internal reconnaissance, brute-forcing and installing webshells, plus one that removed webshells left by rival attackers. CyberScoop reported that Google tied the operators to Russian speakers based on Russian-language comments in the code, including a configuration file that labeled LLM API keys as being for "auto-exploitation (exploit_agent pilot)." Victim telemetry pointed to French websites in media, entertainment and education. Google did not name a group or a state.

Until now, Google says, it had mostly seen adversaries use AI to speed up work that people still directed: writing phishing lures, code and summaries of stolen data. The new report describes the shift past that point, with GTIG noting that open-source agentic pentesting frameworks such as CAIRN and PentAGI now let attackers "automate discrete phases and specific operational tasks during live intrusions."

Why Patches Became the Target List

The choice to mine patches rather than hunt for new flaws is deliberate, and Google explains the logic. "Threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities," the report says, because the research is confined to a finite set of material: source code before and after a fix, including pre-release or preview channels, plus release notes, security write-ups and proof-of-concept code. Hunting zero-days, by contrast, "requires casting a wider net."

Recent cases show how fast that pipeline already moves. Two PaperCut zero-days exploited in late August, an authentication bypass and an unsafe class-loading flaw that chain into unauthenticated remote code execution, had working exploits within a day of disclosure, built with LLM help by security researchers and Huntress. GreyNoise saw a campaign against hundreds of organizations four days after disclosure, with the attacker moving from an initial foothold to domain administrator within minutes. And for two PHP flaws patched on October 21, a proof-of-concept video and early patched code appeared before the official release, letting anyone compare versions and build exploits before most administrators could patch at all.

The volume is climbing too. Google counts overall disclosures and exploits as having nearly doubled between January and August 2026. Exploited zero-days rose more slowly, from an average of 8 a month in 2025 to 11 a month in 2026, before spiking to 22 in August. As The Hacker News noted, LLM-driven pipelines that automate vulnerability analysis and proof-of-concept development shrink the window between patch disclosure and active exploitation. There is a side effect defenders already feel: in the days after the PHP patch, hundreds of claimed proofs of concept were published, most of them non-functional and likely AI-generated, which Google says creates extra verification work for security teams.

The Noise Is the Opening, for Now

There is one advantage on the defending side, and Google is explicit that it is temporary. These frameworks "currently trade stealth for velocity," the report says. Unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers and leaves distinct code-level traces. As Help Net Security summarized, GTIG says the underlying intrusion mechanics have not changed, with compromised credentials and unpatched systems still doing the work, but the speed and scale at which frameworks exploit those weaknesses is new. Google expects less human involvement and more stealth as the tools and models improve, and says actors less concerned with staying hidden, including many cybercrime groups, will keep adopting them. For IT teams, that points to a short list:

Patch management has always been a race. What Google's report changes is the identity of the other runner: not a skilled researcher who needs days to reverse a fix, but software that reads every patch for popular web platforms as it lands and tries the result against live targets. The organizations that keep up will be the ones that stop treating a vendor's fix as the end of the problem and start treating it as the starting gun.

Share

More in AI & Automation

All Resources →